In the risk factors of its revised S-1, Anthropic says its internal estimates and its expectations for future models, rather than its current ones, indicate that it "may imminently meet" its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI." The same filing says: "We have not designated any of our current models as exceeding this threshold."
The coverage calls it RSI, recursive self-improvement, and the headlines went straight to 'self-building AI.' I read that framing as half right. The capability Anthropic describes is real enough to plan around. The statement doesn't describe a model quietly rewriting itself overnight, and the bottlenecks don't work that way either.
I'm working from the filing as TechCrunch quoted it, Anthropic's essay "When AI builds itself," and its published Responsible Scaling Policy. I don't have access to the underlying evaluations, and I'm not going to pretend I do. Where I'm reading between the lines, I'll say so.
RSI as a pipeline: kernels, ablations, evals, and picking the next idea
Building a frontier model is a pipeline of work: writing training code, optimizing GPU kernels, designing and running ablations, cleaning and generating data, building evals, reading results, and deciding what to try next. RSI is the point where a model does enough of that pipeline, well enough, that each generation shortens the time to the next one. The improvement compounds because the thing doing the improving is also the thing being improved.
Each of those is a single link. Anthropic is asking when they connect.
Anthropic says that as of May 2026, more than 80% of the code merged into its codebase was authored by Claude.
Writing most of a lab's code is one thing. Deciding what research the lab should do next is another, and that second one is where 'close' starts to matter.
Anthropic already drew the line it's warning about
This warning didn't come out of nowhere. Version 3.0 of Anthropic's Responsible Scaling Policy, effective February 24, 2026, describes the line as "AI systems that can fully automate, or otherwise dramatically accelerate, the work of large, top-tier teams," and treats it as met when a model could compress two years of 2018 to 2024 AI progress into a single year. That is the definition the filing says future models may imminently meet. It's a big line, not an entry-level one. I still don't read it as a claim that any model is improving itself without supervision. The filing says the opposite: Anthropic doesn't expect to safely use Claude to fully autonomously build successive frontier models without meaningful human supervision, and it is "considering updates to our Responsible Scaling Policy to keep humans in control."
Anthropic's essay puts numbers on the trend: the length of software tasks models can complete "has been doubling roughly every four months, up from an earlier trend of doubling every seven months," and Claude Opus 4.6 managed 12-hour tasks. That covers real, unglamorous research engineering: setting up an ablation, debugging a data loader, porting a kernel. The essay also concedes that "large performance gaps persist when it comes to Claude exercising judgement in choosing goals." That's the judgment a senior researcher brings to a failed run.
GPUs, training curves, graders, and taste: what throttles the loop
The 'self-building' framing assumes intelligence is the bottleneck. In practice, at infrastructure scale the bottleneck is usually the queue: jobs waiting on hardware, on reviews, on results.
GPUs and training curves don't recurse
A model that designs better experiments still needs GPUs to run them. Anthropic's own essay sketches a future where "the pace of progress in AI development becomes determined entirely by the availability of compute." If an automated researcher came up with ten times more good ideas, my bet is they would still compete for the same cluster a human team does. Better ideas raise the value of each GPU hour, and the supply of GPU hours stays where it was.
Much of the signal in ML research only shows up after a run finishes. A model can't think its way past a training curve it hasn't seen yet. I expect the early gains to cluster in work with fast feedback: kernels, data pipelines, infrastructure code, small-scale experiments. Anything where you learn the answer in minutes.
Who grades the successor
A model optimizing against a measurable objective will find ways to satisfy the measurement without doing the work. In a coding benchmark, that costs you a misleading score. In a loop where a model evaluates the research that produces its successor, the eval itself becomes the target: a model that learns which loss curves and benchmark deltas get a run approved can drift toward producing those signals instead of the underlying improvement. You end up with a successor that looks better on every dashboard and isn't, and the next generation inherits whatever blind spot the last one exploited. Humans in the loop slow things down. They also catch exactly this failure, and I'd expect Anthropic to keep them there longer than the raw capability would justify.
Picking the experiment worth a week of cluster time
Choosing which of a hundred plausible experiments deserves a week of cluster time is the scarcest skill in any lab. I don't know how close current models are on that, and I suspect nobody outside Anthropic does.
The S-1 timing, and why I still believe the claim
The cynical read writes itself. A company in registration wants to sound like it's ahead, and 'our models are about to help build themselves' is a strong pitch. AI labs have used capability warnings as marketing for years; the warning and the brochure often read the same.
I don't think that read holds up well here. I'd expect securities lawyers to have read every word of this one. A warning is harder to fake than a demo video, and saying you're near the line commits you to the expensive safeguards that come with crossing it.
My position: the claim is credible in its narrow form. Future models that dramatically speed up a lab's research, with humans still choosing the goals and holding the controls, look plausible from Anthropic's own data. The headline invents the jump from there to an autonomous loop, which the filing says Anthropic doesn't expect to run without meaningful human supervision.
The security problem moves to the weights
The weights question worries me more than the capability claim, partly because access control is a core part of security work. A model that can meaningfully accelerate AI research is a strategic asset in a way a chatbot isn't. An attacker who steals its weights walks away with a research team that works around the clock. That changes who wants the weights and how much they'll spend to get them.
I doubt any lab can fully hold that line against a determined nation-state, and if a breach comes, I'd bet it comes from inside. Insiders, the supply chain, and the people with legitimate keys to the cluster are the realistic attack paths. The same conclusion applies to agents: the threat model is whoever already has access.
There's a stranger version of the insider problem too. A model doing research inside the lab touches the lab's code, its infrastructure, and in some setups its own training pipeline. The obvious control is to treat that model as an untrusted insider: scoped permissions, logged actions, approval gates it can't satisfy itself. I'd like to know how Anthropic structures it.
What changes for people who just buy the API
Most people reading this don't train frontier models. They run workloads on them. RSI talk feels abstract until you look at what faster research cycles do downstream.
- Model churn speeds up. Faster research means releases land closer together and deprecation calendars tighten, so any workload pinned to a specific model version needs a migration plan that assumes shorter notice.
- Your eval suite becomes the most valuable thing you own. Without a way to measure whether a new model does your job better, you can't safely take advantage of faster releases.
- Agent permissions deserve a second look. A more capable model running inside your tenant with standing access is a bigger insider risk, and the risk doesn't shrink because the model got smarter.
- Who signs off when a vendor swaps the model behind an endpoint you depend on? Capability jumps can arrive between contract renewals. Build that review step now.
- Cost assumptions could swing in either direction. Training efficiency sometimes shows up as cheaper inference and sometimes as bigger models at the same price.
Evidence that would confirm or sink 'very close'
- The evals behind the claim. I want Anthropic to publish them, ideally in the AI R&D section of the next system card, in enough detail for outside researchers to argue with.
- Third-party evaluation from METR, the UK AI Security Institute, or US government evaluators. A vendor grading its own model is a starting point.
- Release cadence across the next two or three model generations. If the loop has started closing, the gaps between releases should shrink while the size of each jump holds steady or grows.
- How OpenAI and Google DeepMind respond. Silence would mean they're either behind or reading their own evals differently, and both would tell me something.
Self-building AI is a bad label for what Anthropic described. A better one: a model that does a large share of the grunt work in AI research, inside a lab that still controls compute, evaluation, and the decision to train. That label is also more useful, because it points at the actual control points: compute allocation, verification, and who holds the weights.
The loop doesn't need to close completely to change things. It only needs to shorten. If Anthropic is right about where its future models will sit, the next year of releases will show it, and organizations with solid evals and tight agent permissions will absorb that change far better than the ones still running on defaults. What I can't tell yet is how long the humans stay in the loop once the loop gets fast enough that they start to feel like the queue.
If you work on ML infrastructure and have watched coding agents take over real research engineering tasks, I'd like to know where they still fall over. Is it experiment design, debugging a failed run, or knowing which result to trust?